We will show how to use Instructor with Gemini Flash to get structured outputs with a human in the loop.
Suppose you have an outline for a course you're teaching and you want to break it down into n topics with this structure:
Topic 1
|__ Subtopic 1
|__ Concept 1
|__ Concept 2
| ...
|__ Subtopic 2
|__ Concept 1
|__ Concept 2
|__ Concept 3
| ...
| ...
Topic 2
|__ Subtopic 1
|__ Concept 1
| ...
| ...
.
.
.
Topic n
|__ Subtopic 1
|__ Concept 1
| ...
| ...
Instructor allows us to set Topic as the return type of our LLM call. So we can guarantee that the output will be a Topic object, and not just a string.
We could just run a loop to generate a new topic n times, but what if we‘re not happy with topic 2/n? What if the model comes up with a naming convention that we don’t like and our prompt wasn't that specific in the first place? What if we just want to add some sutble changes to the latest generated topic?
We just wasted all those tokens, minutes, dollars.
That‘s where Burr comes in. By having the creator function as a node, we will ask the user for feedback on the generated topic. If the user is happy with the topic, we will move on to the next one. If not, we will use the feedback to generate the topic again. We will keep doing this until we’ve reached n topics that the user is happy with.
To run this example, you need to install the following dependencies:
pip install -r requirements.txt
Then if you run jupyter lab you can open/execute the notebook. This simulates an interactive web app. The notebook is the best way to get started.